Automatic Epileptic Seizure Onset Detection Using Matching Pursuit: A Case Study Thomas L. Sorensen † , Ulrich L. Olsen † , Isa Conradsen † , Jonas Henriksen † , Troels W. Kjaer * , Carsten E. Thomsen ‡ and Helge B. D. Sorensen † Abstract— An automatic alarm system for detecting epileptic seizure onsets could be of great assistance to patients and medical staff. A novel approach is proposed using the Matching Pursuit algorithm as a feature extractor combined with the Support Vector Machine (SVM) as a classifier for this purpose. The combination of Matching Pursuit and SVM for automatic seizure detection has never been tested before, making this a pilot study. Data from red different patients with 6 to 49 seizures are used to test our model. Three patients are recorded with scalp electroencephalography (sEEG) and three with intracranial electroencephalography (iEEG). A sensitivity of 78-100% and a detection latency of 5-18s has been achieved, while holding the false detection at 0.16-5.31/h. Our results show the potential of Matching Pursuit as a feature extractor for detection of epileptic seizures. I. INTRODUCTION About 1 % of the world’s population suffers from epilepsy [1][2], making it one of the most frequent neurological disorders only outnumbered by stroke and headache [3]. About 75 % of epilepsy patients can be seizure free on antiepileptic drugs, and some of the remaining 25 % can be treated with other procedures, like surgical resection of the epileptic focus, a vagus nerve stimulator or a ketogenic diet [4]. The goal of this study is to build an automatic onset detection for epileptic seizures. Such an alarm would give patients suffering from epilepsy an opportunity to leave their homes knowing that family or medical personnel can come to their rescue if they encounter a seizure. Furthermore it is important to register the number of seizures the patient encounter in a given time frame. This can give medical doctors insight on how well a treatment is working. It can also be important to know when a patient has a seizure, in case of acute treatment, or if a tracer drug has to be administered for an ictal SPECT-scan. An automatic trigger for the vagus nerve stimulator is another possibility, since it has the greatest effect if it is activated early in the seizure [5]. An automated seizure detection system would also assist in † Department of Electrical Engineering, Technical University of Denmark, Kgs. Lyngby, Denmark * Department of Clinical Neurophysiology, Rigshospitalet University Hos- pital, Copenhagen, Denmark ‡ Department of Odontology, University of Copenhagen, Denmark T.L. Sorensen: thomas.lynggaard@gmail.com U.L. Olsen: ulrich.olsen@gmail.com H.B.D. Sorensen: hbs@elektro.dtu.dk detecting seizures in large encephalography (EEG) data sets, that often include recordings from several days. Automatic seizure detection is not a new idea. Through the past couple of decades many attempts have been made, to find the optimal algorithm for classification, primarily using intracranial EEG (iEEG) or scalp EEG (sEEG) [6]. More recently other approaches have been attempted such as accelerometers, electromyography (EMG) and angular velocity recordings [1][4]. We have applied the Matching Pursuit algorithm on both iEEG and sEEG data providing features, which will be used with the Support Vector Machine (SVM) classifier. The algorithm was first used to study ictal EEGs by Jouny et al. in 2003 [7]. However this is the first time SVM has been combined with Matching Pursuit for seizure onset detection. II. METHOD A. Clinical data We have included six patients (pt.) with a total of 133 seizures in 305 hours of recordings (rec.) in this study. To investigate if the robustness of the algorithm depends on whether data is collected intracranially or extracranially, two of the patients are recorded with sEEG and two are recorded with iEEG. The EEG-data is recorded at the Epilepsy Mon- itoring Unit (EMU) at Rigshospitalet University Hospital, Copenhagen. The sEEG-data are recorded at a sampling frequency of 200 Hz from patients admitted for diagnostic workup, using Stellate TM Harmonie with 21-25 EEG chan- nels, placed using the 10-20 system. TABLE I PATIENT INFORMATION Pt. Sex Age Rec. Modality Type # of Seizures P1 M 6 49 h sEEG pGTCS 10 P2 M 63 8h sEEG CPS 49 P3 F 33 44 h sEEG SPS 35 P4 M 45 95 h iEEG CPS 20 P5 F 28 66 h iEEG SPS/CPS 13 P6 M 45 43 h iEEG SPS/CPS 6 Sum 305 h 133 • pGTCS = primary Generalized Tonic Clonic Seizures • CPS = Complex Partial Seizures • SPS = Simple Partial Seizures 32nd Annual International Conference of the IEEE EMBS Buenos Aires, Argentina, August 31 - September 4, 2010 978-1-4244-4124-2/10/$25.00 ©2010 IEEE 3277